Cerebras after the selloff: real demand, unfinished economics

A lower price still asks for a large business

Cerebras closed October 23 at $110.82. That is a calculated 40.1% below its $185 IPO offer price and 64.4% below its first trading-day close. Those are two different reference points: the latter also measures the unwinding of the debut premium. Neither measures the stock's discount to intrinsic value. ^[1][2]

My assessment is that Cerebras has a credible technological advantage and substantial commercial validation, but the shares still require an aggressive operating outcome. The opportunity is to own a valuable supplier of fast AI inference before its capacity becomes fully productive. The trap is to assume that a large contract, a spectacular speed demonstration and a fallen share price already establish attractive returns on capital.

The distinction matters because the company is building more than a chip business. Its cloud ambitions require the silicon to become installed capacity, the capacity to become customer usage, and that usage to produce cash after infrastructure spending. The investment case lives along that chain. ^[10][11]

What the wafer actually changes

The Wafer Scale Engine concentrates processing and high-bandwidth local memory across an unusually large piece of silicon. The prospectus describes WSE-3 with four trillion transistors and 44GB of on-chip SRAM. The architectural objective is to reduce costly movement of information between separate processors and memory. It does not mean every frontier model's weights fit inside one wafer. ^[2]

The current system story has also moved beyond the original WSE-3 pitch. Announced in August, CS-4 combines three WSE-3 Turbo processors with a redesigned Nexus rack. Cerebras advertises inference up to 30 times faster than GPU systems across its selected comparison set; the supporting material includes external comparisons, internal benchmarks and projections. That is a workload-specific performance claim, not a universal economic advantage over every GPU configuration. ^[3]

The strongest commercial application is the part of inference called decode: generating the answer one token after another. Prefill, which processes the incoming prompt and context, has different computational requirements. Cerebras's announced AMD integration assigns prompt processing to AMD hardware and fast token generation to the wafer engine. AWS has separately announced a Trainium–Cerebras approach intended for Bedrock. ^[4][5]

That division supports a more precise investment thesis than wholesale GPU displacement. Cerebras can earn a useful place in a mixed system by doing a valuable stage exceptionally well. It also means its addressable revenue should be assessed against the work it actually performs, rather than against the entire AI infrastructure budget.

Manufacturing remains concentrated: the prospectus identifies TSMC as the foundry for Cerebras's proprietary processor. A distinctive architecture does not remove dependence on fabrication capacity, acceptable yields and timely system delivery. ^[2]

Where future demand could become durable

The attractive demand driver is work that waits on successive model responses. A coding agent may need to generate a change, inspect a result, revise its approach and try again. If the model is on the critical path repeatedly, faster generation can improve the usefulness of the whole application. OpenAI's Codex-Spark deployment and its later Ultrafast preview provide concrete examples of customers pursuing this latency-sensitive category. ^[7][8]

There are three plausible ways that demand expands. Existing applications can pay for a faster service tier. Faster interaction can make previously awkward applications usable. And agents can consume more inference because they perform more steps within an acceptable waiting period. These are mechanisms for growth, not a numerical market forecast.

The limits are equally consequential. Faster output cannot compensate for a wrong answer, an unsupported model or a workflow dominated by slow external tools. OpenAI's own Spark announcement separates output generation, prefill, tool execution and network overhead when discussing task duration. A token-speed advantage therefore cannot simply be multiplied into an equivalent gain in completed work. ^[7]

My demand test is whether a customer receives enough additional useful work to justify the total bill. That bill includes serving infrastructure, energy, idle capacity and integration. Cerebras does not need every application to value latency. It needs a sufficiently large group that values it consistently enough to support attractive pricing and utilization.

Model changes create a moving target. Cerebras's October 22 discussion of mixture-of-experts training explains how external weight streaming addresses memory constraints, while compute utilization still requires software work. This is evidence of continued engineering progress, and a reminder that wafer size alone does not remove every bottleneck. It is a training discussion, so it should not be treated as proof of an across-the-board inference victory. ^[12]

OpenAI: a real deployment, with boundaries

The relationship has developed through several distinct stages:

• January: OpenAI announced 750MW of Cerebras capacity, coming online in tranches through 2028. Its stated strategy was a resilient portfolio matching different systems to different workloads. ^[6]
• February: Codex-Spark launched as a research preview on Cerebras. OpenAI explicitly said GPUs remained foundational and delivered the most cost-effective tokens for broad usage, with Cerebras serving workflows needing extremely low latency. ^[7]
• August: OpenAI previewed GPT-5.6 Sol Ultrafast, powered by Cerebras, at up to 750 output tokens per second. Access was limited to an initial group of API customers; this was not evidence of unrestricted availability across OpenAI's products. ^[8]
• Early October: Sam Altman publicly described Cerebras as a close partner with a deep engagement on speed, responding to speculation about the relationship. His statement reaffirmed the partnership without specifying new volumes or routing commitments. ^[9]

The financial evidence is more useful than reassurance alone. Cerebras recognized $56.8 million of OpenAI revenue in Q2, approximately 31.5% of total GAAP revenue, calculated from the filing. The agreement also gives OpenAI an optional additional 1.25GW beyond the committed capacity. Optional expansion belongs in the upside case; it is not the same as an exercised order. ^[11]

This makes the relationship both an asset and a concentration risk. A sophisticated customer has integrated the technology and begun paying for it. That customer also has reasons to maintain alternative suppliers and optimize its fleet economics.

Investors should avoid drawing either extreme conclusion from a product name or a partnership statement. The cited announcements do not establish that Cerebras will serve every future OpenAI fast tier. Nor does OpenAI's continuing use of GPUs establish that the Cerebras agreement has disappeared. The evidence supports an active, complementary relationship; future allocation and profitability still need to be demonstrated. ^[6][7][8][9]

Capacity announcements are an intermediate milestone

In August, Cerebras reported more than 600MW live or under contract for delivery by the end of 2027. That wording combines operating capacity with future delivery. It is not a claim that the entire amount was already energized and billing customers. The company also said it expected revenue to more than triple in 2027; that remains a management ambition rather than a reported result. ^[10]

The proposed Finland project illustrates the difference between announced scale and an operating business. Its headline capacity is 165MW, with an initial construction phase. A project announcement establishes development intent; revenue depends on subsequent completion and customer delivery. It should not be added to other capacity totals without checking for overlap. ^[14]

For shareholders, the relevant progression is contracted power, completed facilities, installed systems, customer acceptance, billable service and cash collection. Each step can change the timing or economics of the next. A growing backlog can coexist with disappointing near-term returns if infrastructure costs arrive before productive usage.

This is also why heterogeneous partnerships deserve attention. AMD and AWS potentially improve the fit between workloads and hardware and can broaden distribution. They also create integration dependencies. The question for the next reports is how much announced collaboration has become reliable, economically productive service. ^[4][5]

Revenue quality matters as much as revenue growth

Q2 GAAP revenue was $180.1 million, including $126.0 million in cloud and other services, up 281%. Cerebras also reported $209.9 million of non-GAAP core revenue. GAAP gross margin was 14%, against a 41% core gross margin; core operating margin was −16%. The measures exclude different costs and revenue adjustments, including stock compensation, customer-warrant amortization and data-center pass-through items. ^[10]

Core reporting can help isolate the operating model management wants investors to assess. It does not make the excluded economics irrelevant. Equity compensation affects ownership, customer equity incentives help secure business, and a data center still requires funding even when a cost is passed through. Both presentations deserve to be read together.

The company reported $25.4 billion of remaining performance obligations at June 30. Its filing expected approximately 22% to be recognized over the following 24 months, with timing subject to change. ^[10][11]

That supports visibility beyond the current revenue base. It also demonstrates why the whole backlog cannot be treated as next year's sales, immediately available cash or high-margin proprietary compute. A contract's value to shareholders depends on delivery costs, the recognition schedule and the capital required to fulfill it.

The cash cushion buys time, not a guaranteed return

At June 30, unrestricted cash and investments totaled a calculated $7.92 billion, alongside $684.7 million of restricted cash and $918.2 million of working-capital debt. First-half operating cash outflow plus property-and-equipment purchases totaled a calculated $596.4 million. This is a simple cash-flow calculation, not a forecast of recurring annual burn. ^[11]

The financing cushion makes the near-term question different from a business scrambling for immediate survival funding. It gives management room to build. It also belongs in the valuation alongside future cash needs: money reserved for expansion cannot simultaneously be assumed to remain untouched surplus capital.

OpenAI's support has a cost structure. Its working-capital financing is a secured loan, not a grant. Separately, its milestone-based warrant can cover up to 33.45 million Class N shares at a nominal exercise price; that maximum is potential dilution, not a count of all shares already issued. ^[11]

Those arrangements help align a major customer with the deployment. Existing shareholders should still evaluate how much of the commercial upside they retain. Rapid growth can be real while growth per diluted share is less impressive.

Why the shares can fall while the business grows

The price record establishes a severe drawdown. It does not, by itself, establish one cause. Four pressures are relevant to assessing the decline: the valuation embedded at the IPO, uncertainty over the revenue ramp, uncertainty over eventual margins, and the growing supply of shares eligible to trade. These are analytical explanations to test, rather than a claim that any one event accounts for the entire move. ^[1][2][10]

The prospectus contains staged lock-up releases, including up to approximately 19.4 million shares on October 28. The remaining lock-up ends at the earlier of the specified post-Q3-results trading day or 180 days after the May 13 prospectus date. The latter computes to November 9, under that prospectus schedule and subject to applicable restrictions or subsequent changes. ^[2]

Eligibility to sell is not evidence that every holder will sell. Unlocking already-issued shares also does not itself create new shares. It can nevertheless expose the stock to additional supply while investors are still debating the operating ramp.

The practical conclusion is to distinguish price pressure from business impairment. A temporary supply imbalance could create an opportunity if the operating case improves. It offers little protection if the expected earnings power is also being revised down. The old IPO price cannot settle that question.

At $110.82, what is an investor paying for?

Yahoo's quote snapshot puts equity value at approximately $26.3 billion. Against the midpoint of Cerebras's $880–890 million 2026 core revenue guidance, that is a calculated 29.7 times core revenue. This is equity value divided by a non-GAAP revenue measure, not an enterprise-value multiple or an earnings multiple. It should not be compared casually with GAAP-sales ratios for other companies. ^[1][10]

The point is that the drawdown has not removed the need for substantial future growth. As a sensitivity exercise, tripling that core revenue midpoint would produce $2.655 billion and reduce the same equity-value multiple to roughly 9.9 times. That is my arithmetic scenario, not management guidance for that exact metric. It assumes unchanged capitalization and says nothing about the costs or dilution required to get there. ^[1][10]

Even that growth outcome would leave the investment sensitive to sustainable margins. A business earning strong returns on a defensible service can support a very different valuation from one that continually reinvests to keep capacity competitive. Assigning a precise price target before resolving that distinction would create an appearance of precision the current record does not support.

A disciplined buyer therefore needs more than conviction that AI demand will grow. The thesis requires that Cerebras captures a valuable part of that demand, retains pricing power, fills its infrastructure and converts the resulting gross profit into cash per share.

Opportunity or trap: the evidence that would decide it

The bull case is that very fast inference becomes an enduring premium service. Cerebras delivers its capacity, OpenAI expands productive usage, additional customers reduce concentration, and better system design improves utilization. Revenue then grows into the installed cost base while shareholders retain a meaningful share of the upside. The existing deployments make that a credible possibility. ^{[3][6][8][10]}

The middle case is a useful specialist business with slower deployment and mixed profitability. Demand exists, but customers bargain hard, capacity takes time to fill, and the return on new infrastructure remains uncertain. In that outcome, commercial success could still fall short of what the present equity valuation requires.

The trap case is that investors mistake fast tokens for a durable profit pool. Large agreements require expensive infrastructure, competing systems improve, speed premiums narrow and equity incentives absorb part of the upside. Revenue can rise through all of this. The failure would be in shareholder economics rather than in the existence of a working chip.

Cerebras has scheduled Q3 results for November 12 after the market closes. That is the next formal checkpoint; this article uses Q2 financials and the October 23 closing price, not unreported Q3 results. ^[13]

The questions I would carry into that report are:

• How much capacity is energized, accepted and billable, separately from capacity announced or contracted?
• Are OpenAI revenue and broader customer adoption growing without deteriorating contract economics?
• Are GAAP and core gross-profit dollars improving together, with a clear reconciliation?
• What are infrastructure spending, cash collection and diluted ownership doing as the business scales?
• What operational evidence supports management's 2027 growth ambition and the announced partner deployments?

My conclusion is a speculative execution opportunity, with a meaningful value-trap risk. The wafer technology and OpenAI relationship provide substance. The share-price decline improves the entry valuation relative to the IPO, but the current revenue multiple still demands a large and profitable future business. I would want stronger evidence of billable capacity and cash returns before calling it a value purchase. A lower quote is an invitation to redo the underwriting; the operating results must provide the reason to buy.